I have also experienced the same issues of inconsistency and the pure low-effort sort of results, and I don’t exactly know why but I hope they resolve the issue soon.
For context, I use the free version and the GPT 4o works like a charm. Still, as soon as I am done with my trial prompts, the performance goes downhill pretty fast, but I have noticed that Github Copilot’s GPT 4o performs better. More reliable so I rarely use the web version of ChatGPT.
Just my perspective on this